Multimodal Representation Learning via Graph Isomorphism Network for Toxicity Multitask Learning.

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning

Background:

• Compound toxicity assessment is crucial in early drug design.
• Predicting diverse toxic effects presents a significant computational challenge.
• Existing methods struggle with the complexity of compound toxicity tasks.

Purpose of the Study:

• To develop a novel multimodal representation learning model for compound toxicity multitask learning.
• To enhance the accuracy and robustness of toxicity prediction in drug discovery.
• To address the limitations of current computational approaches in predicting compound toxicity.

Main Methods:

• Proposed a multimodal graph isomorphism network (MMGIN) model.
• Employed a two-channel structure to learn fingerprint and molecular graph representations independently.
• Utilized feedforward neural networks for multitask learning, including toxicity and category classification.
• Constructed a new dataset (CTMTL) from the TOXRIC dataset for validation.

Main Results:

• The MMGIN model demonstrated significant advancements over existing machine learning and deep learning models.
• Experimental results on CTMTL and Tox21 datasets confirmed the model's superior predictive capability.
• Ablation studies validated the effectiveness of multimodal representations and multitask learning.

Conclusions:

• The MMGIN model offers a robust and effective approach for compound toxicity prediction.
• Multimodal representation learning significantly improves the accuracy of toxicity assessment.
• This work provides a valuable tool for accelerating early-stage drug design and development.